Computational intelligence methods on biomedical signal analysis and data mining in medical records

• Κύριος συγγραφέας: Vladutu, Liviu-Mihai
• Άλλοι συγγραφείς: Μπεζεριάνος, Αναστάσιος
• Μορφή: Thesis
• Γλώσσα: English
• Έκδοση: 2009
• Θέματα: Artificial neural networks; Statistical learning; Ischemia detection; K-windows clustering; Τεχνητά νευρωνικά δίκτυα; Στατιστική μάθηση; Διάγνωση ισχαιμικού μυοκαρδίου; Αλγόριθμοι για clustering; 610.285

This thesis is centered around the development and application of computationally effective solutions based on artificial neural networks (ANN) for biomedical signal analysis and data mining in medical records. The ultimate goal of this work in the field of Biomedical Engineering is to provide the clinician with the best possible information needed to make an accurate diagnosis (in our case of myocardial ischemia) and to propose advanced mathematical models for recovering the complex dependencies between the variables of a physical process from a set of perturbed observations. After describing some of the types of ANN mainly used in this work, we start designing a model for pattern classification, by constructing several local models, for neighborhoods of the state space. For this task, we use the novel k-windows clustering algorithm, to automatically detect neighborhoods in the state space. This algorithm, with a slight modification (unsupervised k-windows algorithm) has the ability to endogenously determine the number of clusters present in the data set during the clustering process. We used this method together with the other 2 mentioned below (NetSOM and sNet-SOM) for the problem of ischemia detection. Next, we propose the utilization of a statistically extracted distance measure in the context of Generalized Radial Basis Function (GRBF) networks. The main properties of the GRBF networks are retained in a new metric space, called Statistical Distance Metric (SDM). The regularization potential of these networks can be realized with this type of distance. Furthermore, the recent engineering of neural networks offers effective solutions for learning smooth functionals that lie on high dimensional spaces.We tested this solution with an application from bioinformatics, one example from data mining of commercial databases and finally with some examples using medical databases from a Machine Learning Repository. We continue by establishing the network self-organizing map (NetSOM) model, which attempts to generalize the regularization and ordering potential of the basic SOM from the space of vectors to the space of approximating functions. It becomes a device for the ordering of local experts (i.e. independent neural networks) over its lattice of neurons and for their selection and coordination. Finally, an alternative to NetSOM is proposed, which uses unsupervised ordering based on Self-organizing maps (SOM) for the "simple" regions and for the "difficult" ones a two-stage learning process. There are two differences resulted from the comparison with the previous model (NetSOM), one is that we replaced a fixed-size of the SOM with a dinamically expanded map and second, the supervised learning was based this time on Radial Basis Functions (RBF) Networks and Support Vector Machines (SVM). There are two fields in which this tool (called sNet-SOM) was used, namely: ischemia detection and Data Mining.